My first-impression is that this site is vaporware generated by someone prompting an LLM to babble about how an LLM could be secured.
At best, perhaps it's an overengineered way to force humans to rigorously validate every action... which the humans are all very hostile to doing.
> [Architecture page] Reasoning is Read-Only: Input=Data, not Instructions
This doesn't make make sense unless the "reasoning" is regular executable code, and then someone still needs to audit the code whenever it changes. The core LLM algorithms simply don't have room for the distinction.
> Complex workflows broken down into encrypted graph segments
Grandiosity++, where does "the architecture" actually end?
Read enough of EGreg’s posts and you’ll immediately recognize that this site is definitely his but at least LLMs stopped him posting about qbix all the time, a sweet mercy
go back through all of your comments from the last decade, group them by comments that promote/mention one of your projects and comments that do not. I’d ballpark 95% of your comments have been self-aggrandizing self-promotion, it is exhausting to read. I guess today was the day I said something when I saw someone else suffering through another one of your comments. You might want to create healthier technology but you have contributed nothing to HN in almost 20 years, almost impressive enough to be an achievement unto itself. I’d eat my hat if I could find a substantive comment you’ve ever written that wasn’t self-promotion.
No you won’t. You won’t eat your hat — but I’ve written quite a few substantive comments on quite a few subjects that didn’t have to do with self-promotion. The difference is that those comments actually contributed LESS to HN, because they involved random arguments that were ultimately forgotten.
Excuse me, for actually trying to build things instead of just bitching and then let people on HN know about them, or posting things that I believe in. According to you, I guess I should have spent more % of my posts being snarky and attacking things others have built, like our friend here.
I have contributed quite a bit — not i n the form of snark, but in the form of actual SOLUTIONS that I built and put on github as open source. If you don’t want to click or use it or even look at it, that says more about you than me.
And yet your next comment after this was, once again, self promotion, proving my point.
Safebots isn’t a contribution. You have just reinvented tools but bundled it with thousands of words of AI generated slop and terrible AI generated “ideas”. Everyone with Grok invents a new “computing substrate”. Yes, “substrate” is such a Grokism that I can immediately identify you use Grok for “thinking”.
If you spent just a little more time reading and a little less time self-promoting you would know that you are 12 months behind everyone else, human in the loop approval flows for tool execution are standard. The examples of LLMs doing destructive things are because users provided blanket approval.
“I have contributed quite a bit — not i n the form of snark, but in the form of actual SOLUTIONS that I built and put on github as open source.”
Code that nobody has ever used. HN is a discussion forum, you haven’t contributed to discussion. You have always been a slop merchant, quite impressive that you were slopping before LLMs, maybe you are an internet innovator after all.
Pretty sure there is nothing about “complex workflows broken down into encrypted graph segments”. Did you just make that up to add some oomph?
The actual point of the article was: recursive self improvement is scary and upredictable, you can get 99% of the benefit by doing the reliable and predictable thing instead. Did you even read the actual article? I didn’t see anything in your comment about that main thesis which is woven through — well — every paragraph.
Update: oh, I went through your recent comment history and apparently most of it is just snarky criticisms of surface level things. Okay :)
> Pretty sure there is nothing about “complex workflows broken down into encrypted graph segments”.
In the quotes I noted [Architecture Page], which is the extremely prominent blue button/link on the SafeBots homepage. (To a PDF.)
It's on the slide "Scalability: Distributed Reasoning & Encrypted DAGs", number 11 with a black background.
> Did you just make that up
I choose to take this false-accusation as a positive sign that we feel the same way.
You read the text I quoted and (just like me) felt it somehow didn't make much sense... and you felt that very so strongly that you could not believe it even existed!
I understand your predilection for being snarky, but you actually helped.
The front page of safebots.ai contains stuff I put up quickly 6 months ago when I was first making the site. I used Google Notebook LM to create the presentation, based on my notes, since people like “pretty designs”. And it put the word encrypted in there, which was wrong. Yes, the data (and graph) is in fact encrypted at rest, so technically it’s true, but not the main point.
I have since actually built the code, and wrote many articles around it, none of which have the word “encrypted”. Here is a good jumping off point:
I didn’t expect you to ignore the entire article I linked to, go clicking around the website and into a PDF from a completely different page, go to slide 11 and post it. Suffice it to say - your conclusion is totally wrong. But I’m glad you highlighted that specific title. It is time to thoroughly redo that front page - so it is more accurate to what the project has become!
At best, perhaps it's an overengineered way to force humans to rigorously validate every action... which the humans are all very hostile to doing.
> [Architecture page] Reasoning is Read-Only: Input=Data, not Instructions
This doesn't make make sense unless the "reasoning" is regular executable code, and then someone still needs to audit the code whenever it changes. The core LLM algorithms simply don't have room for the distinction.
> Complex workflows broken down into encrypted graph segments
Grandiosity++, where does "the architecture" actually end?